Vocabulary-free Image Classification

In recent years, the development of large-scale vision-language models has fundamentally transformed the paradigm of image classification. Despite these models showcasing impressive zero-shot capabilities, they still assume a predefined class vocabulary at test time. However, this assumption may be impractical in scenarios where the semantic context is unknown and continuously evolving. Vocabulary-free Image Classification (VIC) aims to assign input images to classes within an unconstrained semantic space induced by language, without prior knowledge of the vocabulary, thereby enhancing the flexibility and adaptability of the model, which holds significant application value.